FCE: CyberMake

Ubiquitous Computing In Advanced Manufacturing

White paper submitted to the NSF ENG/CISE Information Infrastructure Technology and Applications Initiative 1994.

Abstract: Inexpensive computing and sensing technology is making it possible to measure and store the history of manufacturing equipment and products, infrastructure components, and medical patients. A challenge for computer science is to make effective use of this flood of data. We will develop and demonstrate algorithms, technology, and paradigms for the use of this information in advanced manufacturing. The scope of manufacturing and design will be extended throughout the lifetime of the product. Information from a product such as a piece of factory automation equipment will be used to develop software upgrades for that individual piece of equipment, as well as to refine the designs of new equipment. We will develop machine learning algorithms to detect patterns, optimize performance, customize product behavior, and predict failures. We will also develop large scale simulations to test and verify these learning algorithms, as well as to allow the construction of virtual factories to evaluate manufacturing process designs. We will develop networking services that allow products to communicate with their factories or maintenance organizations, and allocate processing and sharing of data efficiently.

PI: Christopher G. Atkeson (cga@cc.gatech.edu)

Co-PIs: Ronald C. Arkin (arkin@cc.gatech.edu), Richard M. Fujimoto (fujimoto@cc.gatech.edu), Ashok K. Goel (goel@cc.gatech.edu), Jessica K. Hodgins (jkh@cc.gatech.edu), Amarnath Mukherjee (amarnath@cc.gatech.edu), Ashwin Ram (ashwin@cc.gatech.edu), and Karsten Schwan (schwan@cc.gatech.edu).

Research Plan: The Vision

As the prices of computers and sensors continue to decrease, the opportunity arises to widely distribute embedded computing and sensing in the manufacturing process, at the point of sale, and in the products themselves. A challenge will be to make effective use of this flood of data and allow a manufacturer tight control of design refinements, the manufacturing process, and maintenance of products in the field. We expect each installation of a product to be unique, and the product must optimize its effectiveness in that environment based on data it collects and on the experience of similar products in similar environments. We expect feedback from how a product performs or how it is used to guide rapid redesign, so that a steady stream of improved products with short design times results. We also expect to dramatically decrease the cost of manufacturing equipment and product maintenance and failure by automatically predicting when failures are going to occur and preventing their occurrence. Data driven routine maintenance with constant supervision automatically performed by machine will be much cheaper and more reliable than the current practice of periodically scheduled maintenance and/or human monitoring.

An Initial Application

We intend to first focus on products that are used in manufacturing processes, such as electric motors or automation equipment. An example of such a piece of equipment is a packager that sorts discrete objects into bags or packages. Each package must meet weight requirements, and the incoming objects have a range of weights. The packager must decide which object to route to which unfilled package to minimize the amount of excess weight in the package. If a packager has a mechanical failure, the throughput of the manufacturing process must be decreased and the entire line will not operate at full capacity.

This piece of factory automation is manufactured by another company for a broad range of applications. However, if the packager can customize itself for the particular characteristics of the object it is packaging, it can be much more cost effective. The packager manufacturer may provide no flexibility or a set of simple controls to allow a factory worker to adjust for product mean weight, for example. A packager that can track the characteristics of the process such as the variance of the weights can make much better decisions as to which object to pack into which package.

The packaging equipment needs to monitor itself to predict maintenance needs. For example, the monitoring system of a bearing in the coolant system could detect a change in the pattern of vibration. This change might match a change discovered in other bearings manufactured in the same batch, and indicate an impending failure. Constant monitoring by dedicated equipment will replace periodic monitoring by human technicians due to cost and more importantly, reliability. This provides an opportunity to minimize the costs of actual human involvement in routine equipment maintenance and replacement, by scheduling the maintenance when and where it is needed.

Other Applications

We have discussed instrumenting and networking a variety of products with manufacturers, such as large stationary and mobile diesel engines, automobiles, and home appliances such as heating and cooling systems. In each case the products would include an embedded computer and sensing system. The computer would customize the product for the particular operator or environment, and predict future maintenance requirements. Communication with the factory or maintenance organization would allow sharing of computation between the embedded computer and central computer resources. More importantly, communication over the national network would allow sharing of data. The experiences of similar products could be combined for faster and more effective learning. The additional resources of the central facility would allow verification of new control algorithms developed from the operating data. Manufacturers could essentially redesign products already in the field, as well as refining the designs of future products based on data on how those products are actually used.

Research Plan

1) We will develop machine learning, fault detection, and process optimization algorithms that will implement the behaviors described above. We will need to develop experience with the algorithms, to the extent that we trust the algorithms to do something reasonable under a wide variety of circumstances, with as little human intervention as possible. We therefore need to work with real data and test our control algorithms on real equipment.

2) We will develop sources of actual data. We would like to instrument a set of manufacturing processes and several homes to provide real data to develop our algorithms on. This data stream would be made available on the Internet as a general resource.

3) On-line simulation. We need to understand how to couple product design and manufacturing simulations to real-time information on product use. We need to understand how to make such simulations deal with both real-time data and also offer interactive interfaces to enable human users to play `what if' games, to try different solution approaches, etc.

4) Large scale simulations. We need to simulate realistically sized systems. We will develop large scale simulations to provide a general facility for developers to use to test algorithms. These simulations would be available on the Internet as a general resource. In addition, we need to integrate different aspects of manufacturing design, such as combining fluid flow or combustion models simulating dynamic processes in combustion engines with structural models simulating rigid engine components.

5) Intelligent human interfaces. The best way to understand data is to visualize it, and we will need to develop tools that allow humans to explore the data stream generated by these embedded systems. Prototype interfaces will be made available on the Internet as a general resource.

TECHNOLOGY TRANSFER:

The Principal Investigator is in a proposal to the NIST ATP program from Siemens and IBM on "A Generic Anomaly Detection Technology". Funding from the NSF would complement any NIST funding, which is primarily aimed at the industrial participants. The ATP program also provides a route for technology transfer from our NSF funded research to industry. Furthermore, Georgia Tech is formulating its participation in the "Next Generation Vehicle" program. We feel many of the techniques described in this proposal will be critical components in future automobile manufacturing.

Technology transfer is facilitated by the presence of several on campus centers and programs:

The Manufacturing Research Center (MARC) represents a major commitment to manufacturing related research at both the Institute and State levels. At least 4 Industrial sponsors have donated one million dollars or more to be affiliated with this Center. This unit is housed in its own building and affords numerous opportunities for presentation and interaction with potential industrial collaborators. We have discussed our proposal with the MARC director, and we are exploring the possibility of instrumenting and experimenting with on campus manufacturing processes.

The Materials Handling Research Center consists of approximately 30 member companies and 4 universities. Twice annually this group meets on campus, providing us with the opportunity to disseminate our research results to this community as well as directly receive industrial feedback.

Education:

We feel that the proposed effort would have a tremendous educational impact in addition to affecting the students participating directly in the research. This effort will heighten interest in embedded computing and intelligent systems at Georgia Tech. We expect to have a large number of courses that would explore or use the themes described as examples.

The computer integrated manufacturing (CIMS) program at Georgia Tech is a multi-disciplinary endeavor geared to support manufacturing-related education on campus. Student projects could be readily integrated into the proposed research and awarded credit for their participation.

The CIMS/AT&T intelligent mechatronics laboratory provides excellent resources for hands-on manufacturing education available to the investigators of this proposal. AT&T has donated $225,000 to support this laboratory and is keenly interested in projects such as what we are proposing and how it can impact manufacturing education.

Additionally Georgia Tech has recently won a TRP award in excess of $1 million to further enhance their manufacturing educational laboratory facilities and manufacturing curriculum.

Finally, an Integrated Process and Product Design (IPPD) Lab, heavily interdisciplinary in nature, and being funded by the U.S. Army is being set up to support research in the design of complex intelligent unmanned systems. These resources would likely also be available in some capacity for use within this research.